A glimmer is a small light.
This one runs on your own card.

On 10 August Meta opened the weights of Muse Glimmer: 30 billion parameters, dense, text and images, 128K context, Apache 2.0. The official 4-bit build weighs about 17 GB and is sized for a 24 GB GPU. Today the researcher who shipped it said the Muse Spark 1.2 weights are next.

29.6B params, denseApache 2.0 no usage caps131K context~17 GB at 4-bit233 tok/s on an RTX 5090 with DFlash

DOES IT FIT?

Pick your hardware

model0 GB

Choose a card or a Mac, then a build. The bar pours the model into your memory.

HARDWARE

BUILD

Run it tonight

ollama pull muse-glimmer:30b-q4 && ollama run muse-glimmer:30b-q4

The 4-bit tag is the one most people will actually use. Weights are ungated on Hugging Face; Meta lists Ollama, LM Studio, vLLM, SGLang, llama.cpp and MLX among the runtimes. Check the current tag name in the runtime you use.

WHAT IT IS

30B dense

Distilled from Muse Spark, built for always-on local agents: tool calling, screenshots, long tasks. ViT encoder for images.

LICENSE

Apache 2.0

Commercial use, modification, redistribution. No monthly-active-user thresholds. Meta's first fully open model since Llama wound down.

SPEED

233.4 tok/s

RTX 5090 with the DFlash drafter, per Meta. 74.9 without it. 50.2 on an M5 Max, 37.8 on an M4 Max.

DAY ONE

#1 on HN

1,199 points and 637 comments on Hacker News; roughly 750,000 downloads across official and community repos in the first days.

Receipts drawn from the public posts, tap to enlarge

Where the fees go

100% of ${TICKER} creator fees are routed to @ren_hongyu, Hongyu Ren of Meta Superintelligence Labs, through bankrbot on Robinhood Chain.

He shipped Muse Spark 1.2 and Muse Code, and today he said the Muse Spark 1.2 weights are coming soon. The weights are free. The coin is the tip jar.

contract address prints here at launch